019-wallet-backup-merge.rst (9868B)
1 XX 19: Wallet Backup Merging 2 ############################ 3 4 .. warning:: 5 6 This design document is deprecated. The incremental backup and sync 7 protocol described in `DD 92`_ supersedes it. 8 9 .. _DD 92: https://docs.taler.net/design-documents/092-incremental-backup-sync.html 10 11 Summary 12 ======= 13 14 This design doc discusses considerations for merging wallet backups. 15 16 17 Motivation 18 ========== 19 20 The wallet backup functionality is meant to be used primarily with one device 21 per backup account. Multiple devices sharing one backup account is heavily 22 discouraged, as it can lead to unexpected and unwanted user experiences, such 23 as money suddenly vanishing when it has been spent by another device 24 that shares the same backup account. 25 26 However, there are some situations where more than one device 27 accesses the same backup account. This happens when: 28 29 1. A wallet backup is restored on a new device, but the 30 old device is still active. In this scenario, the devices 31 have different device IDs, but share the same wallet root 32 public key. 33 2. An old wallet backup is taken over by an existing wallet. 34 In this scenario, the devices have different devices IDs and 35 different wallet root public keys. 36 ### CG: This is not exactly more than one device accessing the same backup account! 37 ### CG: Maybe formulate intro differently, to talk about key scenarios that deserve consideration / need to be distinguished? 38 3. A wallet device is copied, for example by restoring the whole 39 device from a device-level backup (not a wallet backup!). 40 In this scenario, the devices have the same device ID 41 and the same wallet root public key. 42 43 44 Requirements 45 ============ 46 47 The backup merging must ensure that: 48 49 * No data that the user wants to keep is lost. 50 * No data resurfaces that the user has previously intentionally deleted. 51 * Conflicts should be resolved automatically wherever possible. 52 * The solution tolerates system clocks not being monotonic. 53 54 55 Proposed Solution 56 ================= 57 58 Stored Information 59 ------------------ 60 61 * Every wallet keeps track of the following data: 62 63 * The current version number (positive integer) 64 * The current wallet root public key (Ed25519 public key) 65 ### CG: public key? Not the private key? What is the private information the wallet usually keeps? 66 * The current device ID (human-readable string) 67 * The status of every backup service account (not defined further here) 68 * The last system time observed on the current device (by device ID). 69 70 * A backup blob stores the following information relevant for backup merging: 71 72 * The backup's version number, equal to the version number of 73 the wallet when the backup was uploaded. 74 * The wallet root public key of the wallet that **owns** the backup account 75 * The device ID of the wallet that **owns** the backup account. 76 77 * Every record and tombstone in the wallet's database and the backup blob keeps 78 track of: 79 80 * The version number at which the entry was created. 81 * A timestamp for the entry, based on enforced monotonic time (per device ID). 82 83 The version number is incremented with every operation that adds an record or 84 tombstone to the wallet's database. 85 ### CG: operation or transaction? I would prefer transaction here. 86 87 88 Resolving Conflicts 89 ------------------- 90 91 This section describes how conflicts are resolved when a wallet (with ``wallet_version``, ``wallet_device_id`` 92 and ``wallet_root_pub``) is merged with a backup (with ``backup_version``, ``backup_device_id``, and 93 ``backup_root_pub``). 94 ### CG: The term 'merged' is something I do not like. Is this during 'backup', 'restore', or 'sync'? 95 ### I suspect these cases need to be distinguished, because the user asking for a 'restore' is 96 ### not creating the same situation than a wallet 'sync'ing during an automated backup, and 97 ### that may again differ from an _initial_ backup (where I guess there are no conflicts, but 98 ### to improve understanding 99 100 * If ``wallet_root_pub != backup_root_pub``: The user is shown a warning "the backup 101 account was written to by another wallet and can't be read by this wallet", and offered a dialog to either: 102 103 a. "Take over" the backup account and migrate it to the existing wallet root public key. 104 A clear warning must be shown that this will kick out the other device currently connected 105 to this account **and** will cause all data from the backup account to be lost. 106 ### CG: Do we even want to allow this? How _can_ this happen exactly? What is the relationship between backup account and root key? 107 ### CG: The private account key is derived from the root public key; 108 ### I do not see us saying anywhere that we would even support 109 ### extracting/exchanging account keys. Hence, I think this basically 110 ### cannot happen: to access the backup, I already must know the root private key. 111 b. Remove the backup account from the list. 112 113 Note that when first adding the backup account via a recovery code, there is a third option: 114 Migrate wallet to the account's wallet root public key. This is **only** possible when 115 scanning the recovery code, as the wallet needs the wallet root secret key to migrate 116 to the account. 117 ### CG: I think this should be the only thing that can possibly happen, by UI/UX. 118 ### Of course _theroretically_ someone could extract ONLY an account-priv and 119 ### use that to download the backup, but then they should just not be able to 120 ### decrypt it. End of story. 121 122 * If ``wallet_root_pub == backup_root_pub`` and ``wallet_device_id != backup_device_id``: The 123 user is shown a warning "two wallet devices are using the same backup account", and given 124 the option of: 125 126 a. Taking over the backup account from the existing device. This will not cause data loss, 127 but the other device (if it still exists!) will stop syncing. 128 b. To "abandon" the current wallet. This (optional, but recommended) will sync the current wallet state 129 with a special marker in the backup blob (so the other wallet continues syncing without 130 having to ask the user), and then delete the database contends and create a new ``wallet_root_pub``. 131 ### CG: I do not think we can 'recommend' option b, because we do not know if the other 132 ### device still exists. So the UI should be neutral here between the two equally valid choices. 133 134 * If ``wallet_root_pub == backup_root_pub`` and ``wallet_device_id == backup_device_id``: 135 136 * If ``wallet_version > backup_version``, do a normal backup 137 cycle (merge backup blob into wallet and upload a new backup). 138 ### CG: We should note that the motivation for a merge arises 139 ### from the 3rd scenario under Motivation: full device backup&recovery. 140 141 * If ``wallet_version <= backup_version``, another wallet with the same 142 root public key must have "tampered" 143 with the wallet's state. Do a normal backup cycle, but consider displaying 144 a warning/notification to the user. 145 ### CG: I think there is no point in distinguishing these two cases; 146 ### in both cases, if the merge is non-trivial, something odd happened. 147 ### Still, I am not sure that a warning/notification is helpful, as 148 ### it is hardly actionable for the user. 149 150 151 Garbage-collecting Tombstones 152 ----------------------------- 153 154 Tombstones should be automatically garbage-collected when the following criteria 155 are both fulfilled: 156 157 * The versions of active backup accounts are all larger than 158 the tombstone's version, and 159 * the tombstone exceeded a threshold age (say, 3 days). 160 161 ### CG: I backup at providers A and B. Make transaction T. Then I remove 162 ### provider A from my provider list. I then delete T. Eventually, I backup 163 ### again at provider B without the expired tombstone. Finally, 164 ### I restore from provider A, and then merge with provider B. 165 ### Here, the 'merge' has to be somehow smart enough to drop 166 ### the deleted data from provider A's backup without the tombstone. 167 ### I think we can safely decide that this is the case because 168 ### backup from A says that it was---at the time---synced with provider B. 169 ### However, this means that we do need to additionally retain the 170 ### historic chain of backup providers and their last merge points/versions! 171 172 173 Q / A 174 ===== 175 176 * Q: Why are version numbers and tombstones necessary in backups? 177 178 * A: When syncing with a backup server that still has an old version 179 (but same device ID and wallet root pub), the tombstones ensure 180 that no old data is re-surfaced that has been deleted in later 181 versions. This can happen in practice even with only one device, 182 namely when a backup provider is unavailable for a long time, 183 but then becomes available again. 184 185 * Q: Why are tombstones only GCed after exceeding an age threshold? 186 187 * A: If we deleted them immediately, this might cause data to resurface 188 if a user temporarily removes and adds a backup account (say by accident) 189 that hasn't been synced in a while. 190 ### CG: See above: the timeout does IMO not help here. 191 ### I think we need to track removed backup accounts 192 ### and the last version that was synced there, 193 ### and then basically determine if a sync/merge-chain exists 194 ### from the (possibly resurfaced) transaction version to 195 ### the current wallet version! 196 197 * Q: Why doesn't the wallet root public key get rotated every time 198 that a wallet backup is restored on a new device? 199 200 * A: Because that would mean that old "paper backups" and Anastasis 201 backups stop working, because they are based on the wallet root key. 202 203 * Q: Why can't the wallet obtain some unique devices identifier to exclude 204 case 3 (same device ID, same wallet root pub)? 205 206 * A: Because we don't have a reliable API for this on many platforms. 207 Even if we had one, we shouldn't rely on it.